10. DMAIC: Improve and Optimize

Introduction and Context

The Improve phase of DMAIC converts verified knowledge about a process into better performance. Define identifies the problem, Measure establishes reliable evidence, and Analyze explains the causes. Improve must then answer a more demanding question: which changes will produce the desired result reliably, economically, and without creating new risks? The phase is therefore not merely a brainstorming exercise. It is a disciplined process of generating alternatives, selecting among them, testing them, optimizing process settings, and preparing the organization for controlled implementation.

In Lean Six Sigma, improvement is judged against the critical-to-quality characteristics identified earlier in the project. A solution is valuable only when it reduces the relevant form of waste or variation and improves customer or business outcomes. A faster process that increases defects, a cheaper material that creates safety risk, or a productivity gain that depends on unsustainable employee effort is not a successful improvement.

The Improve phase commonly uses creativity methods, solution screening, Failure Mode and Effects Analysis (FMEA), economic evaluation, multicriteria decision-making, pilot studies, simulation, Design of Experiments (DOE), and response-surface methodology. These methods are complementary. Creativity expands the solution space; analytical methods reduce uncertainty; decision tools make trade-offs explicit; pilots and experiments provide evidence; and implementation planning controls residual risk.

Learning Objectives

By the end of this chapter, students should be able to:

1. Explain the purpose, logic, and outputs of the Improve phase within DMAIC.

2. Generate solutions using structured creativity and Lean problem-solving methods.

3. Prioritize alternatives using FMEA, cost-benefit analysis, and multicriteria decision-making.

4. Distinguish between a pilot, simulation study, designed experiment, and response-surface study.

5. Design an improvement experiment and interpret its practical implications.

6. Develop a risk-based implementation and control strategy.

7. Evaluate managerial, ethical, technological, and organizational issues associated with improvement projects.

From Root Cause to Improvement Architecture

A validated root cause does not automatically identify the best intervention. The team must determine how the cause can be removed, controlled, or made less influential. Three broad solution architectures are useful.

The first is prevention. Prevention changes the process so that the cause cannot occur or is detected before it produces a defect. Examples include mistake-proofing, automated interlocks, redesigned fixtures, or software validation rules. Prevention is generally stronger than relying on inspection after the event.

The second is reduction or control. Here the cause remains possible, but its frequency or effect is reduced. Examples include standard work, parameter control, preventive maintenance, environmental controls, and operator training. These solutions may be effective but can be vulnerable to gradual deterioration or noncompliance.

The third is containment. Containment prevents defective output from reaching the customer, although it does not remove the underlying cause. Inspection, screening, quarantine, and additional testing are containment measures. They may be necessary during transition, but they should not be mistaken for permanent corrective action.

A mature Improve phase seeks a coherent improvement architecture rather than a single isolated action. For example, a manufacturing project may combine a redesigned component to prevent incorrect assembly, an automated verification step to detect residual errors, and a maintenance standard to preserve equipment capability.

Generating Solutions Without Premature Judgment

Teams often restrict improvement because they evaluate ideas while generating them. Structured creativity separates divergence from convergence. During divergence, the aim is to create a broad range of plausible interventions. During convergence, the alternatives are screened against evidence, feasibility, risk, and value.

Brainstorming is useful when the problem is clearly framed and participants understand the process. It should be supported by facts rather than general opinions. Brainwriting, in which participants write ideas before discussion, reduces the influence of hierarchy and encourages contributions from less vocal members. Affinity mapping groups ideas into themes such as materials, methods, machines, measurement, people, information, and environment.

SCAMPER—substitute, combine, adapt, modify, put to another use, eliminate, and reverse or rearrange—provides prompts for redesign. TRIZ uses recurring inventive principles to overcome technical contradictions. Value engineering asks whether each function is delivered at the necessary performance and cost. Design thinking methods, including observation, rapid prototyping, and user testing, are particularly useful where customer experience or human interaction is central.

Lean methods contribute a different perspective. Elimination of non-value-adding steps, reduction of handoffs, point-of-use storage, visual management, mistake-proofing, workload balancing, and standardized work can produce practical solutions without major capital expenditure. However, teams should avoid applying Lean tools mechanically. A kaizen event that rearranges workstations without understanding demand variation or ergonomic constraints may merely relocate the problem.

A useful solution statement identifies the mechanism of improvement. “Train operators better” is weak because it does not specify the behavior, method, or expected effect. “Introduce a keyed fixture that permits loading in only the correct orientation” identifies a mechanism that can be tested and evaluated.

Screening and Prioritizing Alternatives

After generating ideas, the project team constructs a solution portfolio. Alternatives should be described consistently, including the problem addressed, proposed change, expected benefit, resources, assumptions, implementation time, and possible failure modes. This prevents attractive but vague proposals from receiving disproportionate attention.

FMEA as a Risk-Sensitive Selection Tool

FMEA examines how a process, product, or proposed solution might fail. For each failure mode, the team identifies the effect, cause, existing controls, severity, occurrence, and detection. Traditional FMEA often uses ratings from 1 to 10 and calculates a Risk Priority Number (RPN):

RPN = Severity × Occurrence × Detection.

RPN can help structure discussion, but it should not be treated as a precise probability measure. A failure with very high severity may require action even if its RPN is lower than that of several frequent minor failures. Teams should therefore review severity separately, identify safety and regulatory risks explicitly, and use action priority or risk matrices where appropriate.

In Improve, FMEA is applied not only to the current process but also to the proposed future state. A new automation system may reduce manual errors while introducing software faults, cybersecurity exposure, maintenance dependency, or confusing alarm conditions. The future-state FMEA should be completed before full implementation and revisited after pilot evidence becomes available.

Economic Evaluation

Cost-benefit analysis compares expected benefits with the resources and consequences associated with a solution. Benefits may include defect reduction, capacity release, lower material consumption, reduced warranty cost, shorter lead time, improved safety, and avoided regulatory exposure. Costs include acquisition, installation, training, validation, maintenance, disruption, and disposal.

Simple payback is calculated as:

Payback period = Initial investment / Annual net benefit.

For larger or longer-lived projects, discounted cash flow methods are more appropriate. Net present value (NPV) accounts for the time value of money and can include sensitivity analysis for demand, savings, implementation delay, and operating cost. Intangible benefits should not be ignored, but they should be identified transparently rather than converted into unsupported monetary estimates.

A solution with the highest financial return is not automatically preferable. A lower-return alternative may be selected because it is safer, faster to deploy, more compatible with strategy, or less dependent on scarce expertise.

Multicriteria Decision-Making

Improvement decisions normally involve several criteria. A weighted decision matrix makes these criteria visible. The team defines criteria, assigns weights whose total equals 100 percent, scores each option, and calculates a weighted score. Typical criteria include customer impact, quality effect, investment, speed, technical feasibility, employee acceptance, environmental impact, and implementation risk.

For option j, the weighted score can be represented as:

Sⱼ = Σ(wᵢrᵢⱼ),

where wᵢ is the weight of criterion i and rᵢⱼ is the normalized rating of option j. The method is not mathematically objective merely because it uses numbers. Ratings and weights reflect judgments. Robust practice includes documenting assumptions, testing alternative weights, and identifying criteria that are non-negotiable rather than allowing them to be averaged away.

A decision matrix is most useful when combined with dominance analysis. If one option is cheaper, faster, safer, and more effective than another, the inferior option can be eliminated. Where alternatives perform differently across criteria, the matrix supports a reasoned trade-off.

Piloting, Simulation, and Learning Before Scale

A pilot is a controlled, limited implementation intended to test whether a solution works in a realistic setting. It should have a defined scope, duration, baseline, success criteria, measurement plan, responsible owners, and stop conditions. A pilot is not a full rollout conducted informally. It should specify what will be held constant, what will change, and how results will be compared.

Pilots can reveal practical issues that analytical studies miss: workarounds, training gaps, supplier limitations, bottlenecks shifted to another department, and unintended effects on morale or workload. Nevertheless, a pilot may be affected by novelty, unusually close supervision, or a small sample size. The team should distinguish temporary enthusiasm from stable capability.

Simulation is valuable when physical experimentation is costly, dangerous, or slow. Discrete-event simulation can model queues, resources, cycle times, and flow. Monte Carlo simulation represents uncertainty by repeatedly sampling possible values for variables such as demand, processing time, or defect probability. System-dynamics models are useful for feedback, accumulation, and longer-term policy effects.

Simulation is not evidence that the real process will behave exactly as predicted. Its validity depends on assumptions, data quality, model structure, and verification. A model should be verified—built correctly according to its logic—and validated—adequately representing the real system for the intended decision. Sensitivity analysis should identify which assumptions most influence the outcome.

Designed Experiments: Learning the Process Mechanism

When several controllable factors influence a response, changing one factor at a time is often inefficient and can produce misleading conclusions. Design of Experiments varies factors systematically so that main effects and interactions can be estimated.

A factor is an input deliberately changed by the experimenter; a level is a selected setting of that factor; and a response is the measured output. A full factorial design studies every combination of selected levels. A two-level design is often written as 2ᵏ, where k is the number of factors. Fractional factorial designs reduce the number of runs but may confound effects. Central composite and Box–Behnken designs are commonly used when curvature must be modeled.

A sound DOE begins with a precise objective. The team defines the response and its measurement system, selects controllable factors based on process knowledge, identifies noise factors, chooses realistic operating ranges, randomizes run order where possible, and includes replication or center points when appropriate. Blocking may be used to account for known nuisance variation, such as different batches, shifts, or machines.

The basic model for a two-factor experiment may be expressed as:

y = β₀ + β₁x₁ + β₂x₂ + β₁₂x₁x₂ + ε,

where y is the response, β terms are estimated effects, x terms are coded factor values, and ε represents unexplained variation. The interaction term indicates that the effect of one factor depends on the level of another. This is operationally important: the best temperature may depend on pressure, speed, material moisture, or load.

Statistical significance should be considered alongside practical significance. A very small improvement may be statistically detectable in a large experiment but irrelevant to customers or economics. Conversely, an important practical improvement may fail to reach significance in a small, noisy study. Confidence intervals, residual analysis, effect plots, process capability, and confirmation runs should support interpretation.

Response-Surface Methods and Optimization

When the response changes continuously over a region of operating conditions, response-surface methodology (RSM) provides a structured approach to approximate the response surface and locate desirable settings. A second-order model can represent curvature:

y = β₀ + Σβᵢxᵢ + Σβᵢᵢxᵢ² + ΣΣβᵢⱼxᵢxⱼ + ε.

The objective may be maximization, minimization, target achievement, or simultaneous optimization of several responses. Desirability functions convert different responses into scales from 0 to 1 and combine them into an overall desirability measure. For example, a process may seek low defect rate, low cycle time, and acceptable energy use. The resulting optimum must be confirmed experimentally; a mathematical optimum outside the tested region is not a reliable operating recommendation.

Optimization is not always a single point. A robust operating window may be preferable to a sharp optimum that performs well only under exact conditions. Robust design seeks settings that are relatively insensitive to noise factors. In practice, a slightly less efficient setting may be chosen because it remains capable when ambient temperature, operator, material lot, or equipment condition changes.

Risk-Based Implementation Planning

The transition from tested solution to routine operation is a separate managerial challenge. An implementation plan should identify activities, owners, dependencies, resources, training, communication, validation, measures, and escalation routes. The plan should be staged when uncertainty or consequence is high.

Risk-based planning ranks implementation risks according to potential effect, likelihood, detectability, and controllability. High-risk changes require stronger safeguards, such as dual approval, qualification, process validation, contingency procedures, or phased release. In regulated industries, change control, traceability, documented approval, and formal validation may be mandatory.

The control plan should specify the critical process inputs and outputs, operating limits, measurement frequency, reaction plans, and responsibility. Statistical process control can monitor stability, while capability analysis evaluates whether a stable process meets specification. Control limits must not be confused with specification limits: control limits describe process behavior, whereas specifications describe customer or engineering requirements.

Change management is essential because improvement alters routines, authority, information, and sometimes employment conditions. Employees should be involved in designing the future state, trained before implementation, and given a channel to report hazards or unintended consequences. Incentives should not encourage employees to conceal defects or bypass safeguards in pursuit of productivity targets.

Illustrative Case: Reducing Variation in a Pharmaceutical Packaging Line

Background

The following is an illustrative case designed for learning. A pharmaceutical packaging line experiences frequent rejection of blister packs because the printed batch code is incomplete or unreadable. The defect causes rework, line stoppage, and risk of product release delay.

Situation and Analysis

Measurement and analysis established that the problem was concentrated on one printer model and increased when line speed was raised. The team verified the measurement system and found that inspectors agreed on code legibility. Analysis identified three influential conditions: print-head temperature, line speed, and the distance between the print head and pack surface. Variation in pack positioning also contributed.

The team generated alternatives: replace the printer, redesign the mounting fixture, install an automated vision inspection, reduce line speed, introduce preventive cleaning, or optimize printer settings. Reducing speed alone appeared effective but would reduce capacity. Full printer replacement offered technical benefit but required substantial investment and regulatory requalification.

Prioritization

A future-state FMEA showed that a new fixture could reduce positioning variation but might create a maintenance failure if cleaning access were poor. Vision inspection would improve detection but would not prevent poor printing. A weighted decision matrix gave high weight to defect reduction, validation feasibility, operator safety, and capacity preservation. The fixture redesign combined with optimized settings and automated inspection ranked highest.

DOE and Optimization

A designed experiment varied print-head temperature, line speed, and stand-off distance. The response was the percentage of packs meeting the legibility standard. Results indicated that temperature and distance had significant effects, and that the best temperature depended on line speed. A response-surface study then identified a robust operating region rather than a single setting. Confirmation runs at the proposed settings achieved the target defect level under two material lots and two shifts.

Pilot and Implementation

The solution was piloted on one line for three weeks. The pilot included operator training, a cleaning standard, a revised fixture, vision-system alarms, and a reaction plan. The team monitored defect rate, false-reject rate, downtime, and operator interventions. A false-reject issue was discovered during the first week and corrected by improving camera lighting and inspection thresholds.

After validation, the change was released through formal document control. The control plan specified printer setup checks at start-up, periodic verification of code quality, automated alarm response, and escalation when the process exceeded the action limit. The second line was changed only after the first line demonstrated stable performance.

Outcomes and Learning Points

The case illustrates that optimization required more than selecting a faster or newer machine. The effective solution combined prevention, control, detection, experimentation, and implementation discipline. It also demonstrates why pilot results must include unintended outcomes such as false rejects and maintenance burden. Most importantly, the selected operating region was designed for robustness rather than maximum performance under ideal conditions.

Challenges, Limitations, and Common Errors

The Improve phase is vulnerable to solution bias. Teams may become attached to a favored technology before evidence is collected. Another error is confusing correlation with a mechanism: a variable associated with defects may be a proxy rather than a controllable cause. Solutions should be tested against the verified cause-and-effect logic developed in Analyze.

Overreliance on RPN, payback, or weighted scores can create false precision. These tools support judgment; they do not replace it. Poorly defined criteria, hidden political preferences, double-counting of related criteria, and unsupported savings assumptions can distort decisions.

DOE can fail through inadequate factor ranges, unstable measurement systems, uncontrolled noise, missing randomization, or too few runs. Statistical software cannot repair a badly framed experiment. Likewise, simulation can create persuasive but invalid results when the model is calibrated to exceptional historical data or omits human behavior.

Pilots may fail because success criteria are vague, baseline data are missing, or the pilot is conducted under conditions that cannot be reproduced. Implementation can also produce regression to the mean, temporary improvement, or performance displacement from one metric to another. Balanced measures should therefore include quality, delivery, cost, safety, customer impact, and employee consequences.

Ethical, Governance, and Professional Considerations

Improvement decisions have ethical consequences. A project may affect worker safety, workload, autonomy, privacy, employment, or access to services. Cost reduction should not be achieved by transferring hidden costs to employees, suppliers, communities, or the environment. Automated monitoring should be proportionate, transparent, and governed according to applicable privacy requirements.

Data used in experiments and pilots must be traceable and honestly reported. Selective removal of unfavorable observations, manipulation of specifications, or suppression of adverse events undermines both professional integrity and process safety. In healthcare, pharmaceuticals, aviation, food, and other regulated settings, improvement must comply with validation, documentation, and change-control requirements.

Governance should define who may approve changes, who owns residual risk, how exceptions are handled, and when a project must be stopped. A technically successful solution is not organizationally acceptable if accountability is unclear.

Digital Tools and Emerging Directions

Modern Improve projects increasingly use process-mining platforms, industrial Internet of Things sensors, digital twins, automated data capture, and advanced analytics. Process mining can identify actual paths and rework loops in transactional processes. Digital twins can support scenario testing before physical changes. Machine-learning models may predict defects or maintenance needs, while computer vision can provide real-time detection.

These technologies introduce risks. A predictive model may drift when materials, equipment, or customer behavior changes. Biased training data can produce unequal error rates. Black-box recommendations may be difficult to validate in regulated environments. Cybersecurity weaknesses can create new process failure modes. Technology should therefore be evaluated through the same principles used for other solutions: verified benefit, measurement-system adequacy, explainable risk, human oversight, and controlled deployment.

Emerging practice is moving toward sustainable and resilient improvement. Teams increasingly evaluate carbon, water, material circularity, supply continuity, and recovery from disruption alongside traditional quality, cost, and delivery measures. The objective is not simply to optimize average performance but to design processes that remain capable under changing conditions.

Best Practices for the Improve Phase

Effective organizations generally:

• begin with verified causes and measurable CTQs;
• generate multiple solution mechanisms before selecting one;
• involve process owners, operators, customers, technical specialists, and risk functions;
• evaluate the future state with FMEA and explicit economic assumptions;
• use pilots, simulation, and DOE according to the uncertainty and consequence of the decision;
• distinguish statistical significance from practical and financial significance;
• optimize for a robust operating window rather than a fragile ideal point;
• define stop criteria, reaction plans, and ownership before implementation;
• validate benefits after novelty effects have diminished; and
• transfer learning into standard work, training, control plans, and subsequent improvement cycles.

Chapter Summary

The Improve phase transforms analytical understanding into tested and optimized change. It begins by distinguishing prevention, reduction or control, and containment, then uses structured creativity to develop alternative solution mechanisms. FMEA exposes risks in the proposed future state; cost-benefit analysis and multicriteria decision-making make economic and strategic trade-offs explicit. Pilots and simulation allow learning before scale, while DOE identifies influential factors and interactions. Response-surface methods support optimization and robust operating windows. Risk-based implementation planning, change management, control plans, governance, and ethical review ensure that improvement is sustained without creating unacceptable consequences. Digital tools can strengthen experimentation and prediction, but they also require validation, cybersecurity, transparency, and human oversight.

Key Terms

Failure Mode and Effects Analysis (FMEA): A structured method for identifying failure modes, effects, causes, controls, and risks.

Risk Priority Number (RPN): The product of severity, occurrence, and detection ratings in traditional FMEA.

Cost-benefit analysis: Comparison of expected benefits with the costs and consequences of an alternative.

Net present value (NPV): The present value of future cash flows minus the initial investment.

Multicriteria decision matrix: A method for comparing alternatives across weighted criteria.

Pilot: A limited, controlled implementation used to test a proposed solution.

Simulation: Representation of a real system for evaluating behavior under specified assumptions.

Design of Experiments (DOE): Planned variation of factors to estimate their effects on responses.

Factor: An input variable deliberately varied in an experiment.

Level: A selected value or condition of an experimental factor.

Response: The measured output used to evaluate an experiment.

Interaction: A condition in which the effect of one factor depends on another factor.

Randomization: Assignment or sequencing used to reduce systematic experimental bias.

Blocking: Grouping experimental runs to account for known nuisance variation.

Response-surface methodology: Experimental and statistical methods for modeling and optimizing responses.

Robust design: Selection of settings that perform consistently despite noise factors.

Confirmation run: A test conducted to verify predicted performance at selected settings.

Control plan: A documented specification of controls, measures, frequencies, owners, and reaction actions.

Residual risk: Risk remaining after improvement controls are implemented.

Change control: Formal governance of proposed, reviewed, approved, and documented process changes.

Review Questions

Short Answer Questions

1. Why is containment generally considered weaker than prevention in an improvement strategy?

2. What is the purpose of a future-state FMEA?

3. Distinguish statistical significance from practical significance in DOE.

4. What is the difference between control limits and specification limits?

5. Why should an operating window sometimes be preferred to a single optimized setting?

Long Answer Questions

1. Explain how creativity methods, FMEA, economic evaluation, and multicriteria decision-making can be integrated to select an improvement solution.

2. Describe the design and interpretation of a DOE for a process affected by several controllable factors.

3. Discuss the role of response-surface methodology in achieving robust process optimization.

4. Explain how a pilot should be planned, measured, governed, and evaluated before full implementation.

5. Analyze the ethical, organizational, and governance issues that may arise when digital automation is introduced during Improve.

Analytical / Application Questions

1. A process team has three alternatives: additional inspection, fixture redesign, and equipment replacement. Develop a decision framework that considers defect reduction, investment, speed, safety, and residual risk. Explain how you would avoid false precision in the final choice.

2. A DOE indicates that factor A is significant only when factor B is at its high level. Interpret this interaction and explain how it should influence the recommended process settings.

3. A pilot reduces defects but increases operator workload and false alarms. Decide whether to proceed, redesign, or stop the implementation, and specify the evidence and governance criteria you would use.

22. Multiple Choice Questions with Answers

Question 1

Which solution architecture most directly removes the opportunity for a known error to occur?

A. Containment inspection

B. Prevention through mistake-proofing

C. Post-process rework

D. Increased reporting frequency

Correct Answer: B. Prevention through mistake-proofing

Question 2

Why should traditional RPN not be used as the sole basis for FMEA decisions?

A. It cannot include any numerical ratings

B. It ignores all detection controls

C. It may obscure very high-severity risks and treats ratings as overly precise

D. It is applicable only to financial projects

Correct Answer: C. It may obscure very high-severity risks and treats ratings as overly precise

Question 3

A weighted decision matrix is most useful when a team must:

A. Prove that one option has no uncertainty

B. Compare alternatives across several explicit criteria

C. Replace all customer requirements with cost

D. Calculate process capability directly

Correct Answer: B. Compare alternatives across several explicit criteria

Question 4

What is the principal purpose of a pilot in the Improve phase?

A. To bypass formal approval for a permanent change

B. To test a solution under controlled, realistic conditions before scaling

C. To eliminate the need for measurement

D. To guarantee that all future variation is removed

Correct Answer: B. To test a solution under controlled, realistic conditions before scaling

Question 5

In DOE, an interaction means that:

A. Two responses are measured with identical instruments

B. The effect of one factor depends on the level of another factor

C. The experiment has too many observations

D. A factor has no measurable effect

Correct Answer: B. The effect of one factor depends on the level of another factor

Question 6

Which practice strengthens the validity of a designed experiment?

A. Changing several factors informally between runs

B. Using a validated measurement system and randomizing run order where possible

C. Selecting only results that support the preferred solution

D. Avoiding replication in all circumstances

Correct Answer: B. Using a validated measurement system and randomizing run order where possible

Question 7

Why may a robust operating window be preferable to a mathematical optimum?

A. It always has the lowest investment cost

B. It performs consistently despite normal noise and process variation

C. It eliminates the need for confirmation runs

D. It guarantees zero defects

Correct Answer: B. It performs consistently despite normal noise and process variation

Question 8

Which statement correctly distinguishes simulation verification from validation?

A. Verification asks whether the model represents reality; validation asks whether code was typed correctly

B. Verification checks whether the model was built according to its logic; validation checks whether it adequately represents the real system for its purpose

C. They are identical activities

D. Validation is unnecessary when historical data exist

Correct Answer: B. Verification checks whether the model was built according to its logic; validation checks whether it adequately represents the real system for its purpose

Question 9

A pilot improves defect rate but creates excessive false rejects. What is the most appropriate immediate response?

A. Roll out the solution without recording false rejects

B. Investigate and correct the new failure mode before scaling

C. Remove the inspection measure entirely

D. Change the specification to make the result appear successful

Correct Answer: B. Investigate and correct the new failure mode before scaling

Question 10

Which implementation practice best reflects risk-based planning?

A. Applying identical approval and validation requirements to every change

B. Allowing the project champion to own all residual risk informally

C. Using stronger safeguards, validation, and phased release for high-consequence changes

D. Delaying all changes until every uncertainty is eliminated

Correct Answer: C. Using stronger safeguards, validation, and phased release for high-consequence changes